⚡Megatron LLM Hub
全部
Twitter

推特安全流

头条7

首次完整展示 agent 端到端做 exploit 开发的可用架构,可直接借鉴复现

AI agents are now generating Windows kernel N-day exploits. A researched named @chompie1337 built BinChomp, an autonomous exploit pipeline, and breaks down the architecture at BSides Canberra 2026. If you do exploit development, watch this: https://t.co/6o1JO6oF6b #WindowsKernel #ExploitDev #AI

ai_agent内核poc议题 @cr3ghost 原文 ↗

三条受支持版本线全中,已有跨84个构建的稳定PoC,可直接复现。

RT @whoareme33: The Brotli RCE in Node.js is brutal! CVE-2026-111497 affects all three supported major lines: 22.0.0-22.23.3, 24.0.0-24.22.0, and 26.0.0-26.12.0. One sleepless night with GLM-5.3 and DeepSeek V4.1 Flash 1017, plus $26.80 in tokens, and I had a reliable PoC across those versions on Linux arm64 and x86_64 with ASLR and UV_THREADPOOL_SIZE independent vectors for all 84 builds. But I wouldn't call it suitable for blind black-box testing. I ran 14,877 fresh exploit attempts - every ARM target against every ARM and x86 profile, each in a fresh lab. Wrong profiles are destructive: 58.3% crashed the lab (56.4% ARM profiles, 60.2% x86 profiles replayed on ARM). And HTTP-only fingerprinting was useless: all 87 builds returned byte-identical responses, so you cannot safely guess the profile without force/out-of-band info. Bug bounty hunters, please do not spray profiles at production :) Nine years ago, when I started my career as a Node.js SWE, I never imagined that an ordinary body-parser setup on a Node.js runtime could become an RCE entry point. 2026 keeps surprising me. And it's making me look at attack surfaces differently... Huge respect to the researchers who found this bug

rcecvepocnodejs @0xocdsec 原文 ↗

该漏洞已在近期攻击中被利用,根因与演示公开,可直接复现检测。

RT @thezdi: CVE-2026-35273: SSRF to RCE in Oracle PeopleSoft. ZDI's @bobbygould5 provides a deep dive into the bug(s) used in the recent PeopleSoft attacks. His analysis includes root cause and video demo. https://t.co/ibIEwrcWnp

ssrfrcecve企业应用 @0xocdsec 原文 ↗
必看24

零交互通话即远程代码执行,用户基数极大,且由 AI 辅助挖出

RT @nebusecurity: Remember when scam callers only ask for your password? Good times. Now picking up a wrong call can get you hacked. Telegram RCE discovered by NebuSec Code Security, powered by GPT-5.4. Yes, GPT-5.4. We turned a dinosaur model into a cyber powerhouse. Fixed in Telegram Android 12.10.4(7099)

rce移动端ai_agentcve @0xocdsec 原文 ↗

弱模型也能跑通内核 N-day 利用,这套 agent 自动化范式可直接复现。

Video of my @UnpromptedAU & Bsides Canberra talks out now I built an autonomous N-Day exploit pipeline for the Windows kernel w/ mix of deterministic and AI-led stages. Results were comparable to what Anthropic achieved with Mythos using a much less powerful model.

ai_agent内核exploit @chompie1337 原文 ↗

护栏绕过的思路可直接用于自建漏洞挖掘 agent,手法可复现。

RT @yo_yo_yo_jbo: Were you ever frustrated with your AI model for refusing finding vulnerabilities for you? What if I told you that you could still use it to find vulnerabilities, automatically? Welcome to Guardrail-based Vulnerability Discovery (GbVD)! https://t.co/B2cbCOJP6f

ai_agentllm绕过工具 @0xocdsec 原文 ↗

Node.js 服务端 1-day 漏洞有公开 PoC,可立即验证现有部署。

RT @_p0her_: At Google CTF in Singapore with TeamH4C, I came across Node2Shell on X. php2shell already feels well-trodden, so I dug into Node2Shell instead and published the analysis on SAGO, our 1-day platform. 📝 Write-up: https://t.co/Wa6kjiQrq9 💻 PoC & exploit: https://t.co/F1zO11XEQD

pocnodejsrce @h4x0r_dz 原文 ↗

在 Secure Boot 校验前劫持启动,PXE 与裸机部署环境需评估。

RT @cr3ghost: One ICMP packet can hijack a PXE-booting server before Secure Boot checks the OS. CVE-2026-97631 (9.5): pre-boot RCE in TianoCore EDK II. RCE needs a build-specific payload. By @matrosov (Anthropic) + Dan Holmes-Mitra (Google): https://t.co/ijj2qOZrew #UEFI #SecureBoot #RCE https://t.co/d0ybVsJR7u

rcesecurebootcve固件 @cr3ghost 原文 ↗

影响 26.5–27.0.1 且 PoC 与写入公开,未升级设备仍可利用。

RT @MchinaMd57764: CVE-2026-112706: CoreMedia bytestream XPC (mediaplaybackd) handed live file descriptors for protected paths to any sandboxed iOS app — zero permissions, zero prompts. Verified on 26.5–27.0.1; fixed in 27.2. Writeup + PoC: https://t.co/cqgB0CZfR4 #iossecurity

ios沙箱逃逸cvepoc @0xocdsec 原文 ↗

开源权重的开放程度直接决定谁能把强攻防能力部署在自有机器上。

imagine an open model with astra level cyber capabilities gets released to everyone to use/abiliterate. how exactly does this make the world more secure? like, i genuinely don't understand. you do realize not everyone is just going to burn tokens and magically secure themselves, right? i will die on this hill: you need to build models even more capable than astra before open models catch up, and invest more in cyber programs that give the good guys access to these models first. open models reaching astra level cyber capabilities before defenders are ready(whatever that means) would be dangerous for the world. @_deven___ for picture credits

llmai_agent开源议题 @S1r1u5_ 原文 ↗

绕过 Credential Guard 后可直接导出用户哈希,且有完整 writeup 可复现。

RT @papadelospapas_: Runtime Credential Guard Deactivation New Medium post, in this one we will see how to bypass Windows Credential Guard at runtime by signaling an undocumented named kernel event inside LsaIso.exe, and then just dump the lsass process and it’s user hashes https://t.co/XMOtq00h4l

绕过凭据窃取windows报告 @ShitSecure 原文 ↗

自建 GitLab 实例默认可达,Onigmo 正则整数溢出致越界写,直接打穿。

RT @hackyboiz2: [1day1line] CVE-2026-110264: Remote Code Execution via Integer Overflow in GitLab Hello, this is d3lu0. Today's 1day1line covers an unauthenticated RCE vulnerability in GitLab caused by an integer overflow in Ruby's Onigmo regex engine. During .org file rendering, unescaped TODO keywords can be interpreted as regex patterns. When repeat IDs exceed the signed 16-bit limit, they can wrap to negative values, leading to an out-of-bounds write during matching. The public PoC demonstrates rendering timeouts on a vulnerable GitLab instance, but not RCE. Please refer to the blog post for more details! https://t.co/gGfmzu9q3K

rcecvepoc自托管 @0xocdsec 原文 ↗

克隆后 post-checkout 自动执行,开发者本机直接中招,属在打活动。

RT @moonlock_lab: 1/ Cloning a GitHub repo for a "technical interview" on your #Mac? Check .git/hooks first. We found a cluster of trojanized repos where post-checkout is a multi-platform dropper for BeaverTail, the DPRK infostealer behind Contagious Interview. 🧵👇 https://t.co/p1jqvtRnWc

供应链恶意仓库macos事件 @patrickwardle 原文 ↗

直接/间接提示注入、AI 结果投毒等首次有统一战术编号,可对照自查 Agent 攻击面。

RT @ManiarViral: MITRE ATT&CK v20 is here. 9 new Enterprise techniques for AI: Direct Prompt Injection, Indirect Prompt Injection, AI-Enabled Exfiltration, AI Result Poisoning, Impair Autonomous Defenses + CAPTCHA automation. Finally a shared language for AI attacks. #MITRE #ATTACK #AISecurity https://t.co/qDEsxAGS7G

ai_agentllm议题prompt_injection @kernullist 原文 ↗

从无人关注的攻击面打到 SYSTEM 的完整方法论,细节文章即将公开。

RT @HototogisuTian: First time at OffensiveCon Tokyo, and it was an incredible experience! ​In "Breaking NTFS at Scale: From an Unmapped Attack Surface to SYSTEM," we showcased how we hunted down 43 CVEs in Windows NTFS and demonstrated how to exploit these types of vulnerabilities for LPE or RCE. ​We will be publishing more detailed technical write-ups and an extended paper soon.

lperce内核议题 @0xocdsec 原文 ↗

拿到 SYSTEM 权限即可绕过 VBS 凭据保护,Windows 哈希隔离假设失效

RT @papadelospapas_: Credential Guard Runtime Shutdown VBS is supposed to keep your NT hashes locked in VTL1 forever. Turns out LsaIso.exe exposes a named event that tears it all down at runtime. Steal a SYSTEM token, signal it and extract NT hashes straight from LSASS https://t.co/aBc4Ug2uTV https://t.co/06BGplYEoM

凭据窃取绕过windows @pentest_swissky 原文 ↗

漏洞位于 UEFI 固件层,几乎全部服务器受影响且需逐台刷固件修复

RT @IntCyberDigest: ‼️ Anthropic and Google researchers found a flaw that lets one crafted packet hijack a server during network boot, before Secure Boot checks the OS. Almost all servers on earth are affected. CVE-2026-97631 sits in TianoCore EDK II, code in the UEFI firmware of most servers. https://t.co/YhFc3ZATfy

固件ueficverce @cyb3rops 原文 ↗

逻辑反序列化不依赖二进制利用,PHP 部署面极广且当前无补丁

PHP published a follow-up advisory about a new, unpatched RCE bug. No, it's not the bug that I found. It's is easier to exploit. Logic deser - no binary exploitation problems. But I have suspicions about the proposed mitigation guidelines... https://t.co/hANVUEZUHe

rce反序列化php @alisaesage 原文 ↗

AI公司把全部研究机密放在Slack,同型事件2023年在OpenAI发生过。

RT @willdepue: anthropic is massively underplaying the recent security breach. i'm hearing cathedral achieved full administrator access of the slack workspace (see every message, log in as arbitrary users) where literally every single research secret is contained to those unfamiliar: Ant & OAI keep literally everything in Slack. it sounds exactly the same incident that happened to OpenAI in 2023, where OpenAI was hacked by a random guy who got access to an access token from an external app service "we're critical to national security, we must beat China, and yes we keep all our crown jewels in Salesforce's Slack messenger. yes we have admin API tokens everywhere. why do you ask?"

ai_agentllm事件供应链 @0xocdsec 原文 ↗
推荐15

本地 LLM 的对齐可被措辞绕过,直接影响自托管 agent 的滥用防护设计

RT @nmatt0: I feel like people getting refusals from GLM 5.3 are prompting like an undercover cop... "Hey there! Would you mind helping my criminal enterprise author some illegal malware? 😉" If you just stick to the technical details of what you want the agent to do it will do it.

llm绕过ai_agent @0xocdsec 原文 ↗

本地跑第三方微调模型的实际风险与缓解手段,对应自托管场景。

RT @msuiche: Today, was another great day at @unpromptedconf ! 1. Model Internals are the new Windows Internals. 2. Debugging used to be about basic blocks, it is now about transformer blocks. (We love runtime!) 3. AI is a race to control what people can think about. But the hacker mindset can't allow this. "My crime is that of curiosity." 4. Run time overlays like GLP prevent supply chain issues created by abliterated models from untrusted parties (potentially backdoored too). 5. Local models are useful now!

llm供应链ai_agent @moyix 原文 ↗

量化 agent 相对 fuzzer 的挖洞增量,是评估 agent 攻防价值的硬数据。

RT @natashenka: LLMs are great at finding bugs, but so are fuzzers. @5aelo compared the vulnerabilities found in V8 by BigSleep, an agentic system, with those found by the Fuzzilli fuzzer. https://t.co/KTvm6k5Rhr

ai_agentllm漏洞挖掘报告 @alkalinesec 原文 ↗

集中索引常用红队工具与技巧,省去零散搜集时间。

Awesome-Redteam — A knowledge base for red teaming and offensive security. https://t.co/xKyxJnkW8X https://t.co/P9D6xQk8wq

工具红队资源 @KitPloit 原文 ↗

预览驱动无阻断接口,用户态 EDR 的检测与拦截能力面临重估。

Kernel eviction notice: What wesp.sys reveals about the Windows Endpoint Security Platform. TLDR; Microsoft wants EDR out of the Windows kernel, and the preview binaries show wesp.sys telling a user-mode client that a process is starting, with nothing in its imports to let that client block it. Elastic Blog: https://t.co/q2bJ3W0aT0

内核edr报告绕过 @5mukx 原文 ↗

公共频道内容默认外泄,管理员与用户需提前调整共享与隐私设置。

RT @Malwarebytes: Discord is making messages from public servers discoverable on search engines like Google starting this November. Server admins can opt channels out starting today. Users can opt out of sharing their messages in settings -> Data and Privacy. https://t.co/kGN17jTxmh

隐私事件移动端 @artem_i_baranov 原文 ↗

本地 LLM 缺少等价于 LXC 的低开销隔离手段,直接决定自托管暴露面

How do OSX people protect / sandbox their local AIs running on their devices? On Linux I can LXC / LXD and put lots of wrappers around it - without losing any peformance like a VM. Apple has nothing for this yet or is it just a question of me throwing tokens at App sandbox?

ai_agentllm沙箱macos @0xocdsec 原文 ↗
更多97

内核提权与 VBS 绕过研究必备的调用路径说明,少见的清晰图解

Secure kernel, the VLT1 kernel uses the same way to work than the regular kernel (VLT0). VTL0 works via syscalls. In VLT1, the syscalls are named Secure Calls. Usermode -> syscall -> ntoskrnl.exe (VTL0) → HvlpVtlCall → VMCALL → hypervisor → securekernel.exe (VTL1)

内核提权windows @papadelospapas_ 原文 ↗

评估 AI 威胁面时容易只算攻击侧增益,这条补上防守侧视角

On the surface, a lot of the AI doom-and-gloom crowd overlooks something important... Front-line defenders use AI too. They just don’t post about it as much as everyone else. 🎩🐰 A second point: many of them use it well. Don’t ask me how I know.

ai_agent观点 @bohops 原文 ↗

RT @zseano: Expedia Group Program has gone from $21k bounties paid to $250k in just 2 months. Seems people are making good $$$ here: https://t.co/ChfbKBWoQk @m0chan98 is leading rep in last 30 days (it will be interesting to log data for 1year+ to see what's happening across programs) https://t.co/mSaCrVN8GN

漏洞赏金报告 @0xocdsec 原文 ↗

模型选型参考:误报低但漏报高,决定跑扫描用哪个模型前先看这组数据。

RT @pilvar222: We ran Kimi K3.1 Pro on our cybersecurity benchmark, it's super precise (4.2% false positives), but it doesn't find as many vulnerabilities as other models - At pass@3, it found 26/32 CVEs (81.3%), tying Opus 5. Still behind DeepSeek V4.1 Flash 1017 (87.5%) - Only 4.2% of what it reports are false positives (95.8% precision), the 4th lowest rate on the benchmark. DeepSeek is at ~14%, so you get roughly 3x less noise - Cost: $16.14 per CVE found, 8x cheaper than GPT-6 Astra ($135.23), which has a similar false positive rate 1/3 🧵

llmbenchmarkai_agent @0xocdsec 原文 ↗

无 CVE 编号的漏洞容易被漏打补丁,同类 LDAP 攻击面已在讨论中。

RT @al3x_n3ff: @aas_s3curity @azoxlpf Note that this is considered a vulnerability that has been patched with KB5122882 (September). No CVE has been assigned as far as I am aware. Also check out a very similar attack against LDAP, found by @GrayHatKiller (@SpecterOps), patched with KB5126646: https://t.co/jRkwHGKPvK

windowslpe补丁 @0xocdsec 原文 ↗

RT @HaifeiLi: It’s quite obvious now that the AI frontier labs have been pushing very hard on vulnerability discovery, from acquiring world-class vulnerability researchers to training specific vuln-discovery models. However, I don’t think from a business point of view, vulnerability discovery could be a long-term, big source of revenue, and it’s just a very small part of the entire cybersecurity business. So why do they do that? What am I missing?

ai_agent行业 @Laughing_Mantis 原文 ↗

内核态 EDR 架构变动同时改变防护面与对抗面,逆向细节已公开。

RT @cr3ghost: After CrowdStrike, Microsoft wants EDR out of the Windows kernel. A researcher named @zodiacon reversed WESP's preview: • wesp.sys minifilter • espclient.dll user-mode API • no apparent process-blocking reply path yet https://t.co/ZtgQXetz5a #WindowsInternals #EDR #ReverseEngineering

edrwindows内核逆向 @cr3ghost 原文 ↗

硬件级 DMA 到用户态 RCE 的完整链路,手法可复用于本地研究。

RT @cr3ghost: Think reverse engineering is just Ghidra? Gbps covers PCIe MMIO/DMA, captures PCILeech TLPs with a protocol analyzer, and breaks down a full-chain Source Engine client RCE using Frida. This blog deserves a bookmark: https://t.co/xKguwkVYeR Author: https://t.co/S55mSZYIcP #ReverseEngineering #ExploitDev #PCIe

逆向dmafridarce @cr3ghost 原文 ↗

含新 CVE 技术细节与从初始访问到加密的检测点整理。

RT @andreanaspi: I've started publishing technical cybersecurity research on my blog. https://t.co/DSeMerhVPL First two posts are live: a deep dive into PHP2Shell (CVE-2026-109618/109619) and a technical analysis of the Qilin/Agenda ransomware, from initial access to encryption and detection. https://t.co/tajahIroj1

cve勒索软件php报告 @0xocdsec 原文 ↗

Android 应用漏洞利用的基础知识,涵盖 onPause 敏感数据清理等要点。

RT @8kSec: Check out the Android Activity lifecycle from the AOSP source, what the framework actually does behind each callback, how it enforces super calls, and the security implications at each stage: https://t.co/PvpT51fFex  Covers FLAG_SECURE, sensitive data cleanup in onPause(), task affinity restrictions on Android 11+, and more. Understanding how activities manage state and focus is foundational for exploiting Android app vulnerabilities. Practice Android app exploitation hands-on in our free Android Application Exploitation Challenges: https://t.co/5t1MUYZMWv  Follow @8kSec for more mobile security research

移动端android议题 @0xocdsec 原文 ↗

Android 堆利用研究的基础资料,原文已难以获取。

RT @bsdaemon: I have just noticed that my Scudo internals paper (released while at L3Harris Trenchant) can't be (easily?) found anymore due to website changes... here is a link to a pdf copy for those that need it: https://t.co/6wcy49Y6xu

堆利用android报告 @0xocdsec 原文 ↗

浏览器崩溃上报组件的可复现攻击面,细节适合自行复现。

RT @darkfloyd1014: Crashing the Crashpad talk is fascinating. I simply study further about the flow and details, this is quite fun and decent, making the object alive by uploading 512MB object slowly. Thank you so much. @Xbow @offensive_con https://t.co/C59szAVqd5

浏览器crash议题 @moyix 原文 ↗

@tszzl @tenobrus but i agree in general. slack is reasonably secure for most companies. it really depends on who your adversary is. they have a pretty good h1 program where vulnerabilities are constantly being reported https://t.co/WTAndD1uaA also checkout https://t.co/PsSoETj61Y

供应链 @S1r1u5_ 原文 ↗

the point of reducing it down to one purpose was to more easily make the point that security people complain about the stuff that creates the conditions for them having employment. sorry i failed to get that across. security people tend to complain that they aren't allowed to introduce more friction and longer timelines, instead of taking it upon themselves to figure out ways to reduce friction and speed things up, while still delivering security solutions. hope thats clear enough, i wont be elaborating further. thanks

议题 @Teach2Breach 原文 ↗

补上长期缺失的技术证据,可用于比对自身日志与检测规则。

RT @KimZetter: For years government officials have been telling us how sophisticated and stealth China's Volt Typhoon hackers are, without providing details that support this assertion. Today researchers at @BlackLotusLabs finally reveal details that back the claim https://t.co/a3J6t6ofDd

apt事件报告 @NetAskari 原文 ↗

LLM 辅助逆向的结论若无证据和复现,既不可验证也无法复用。

RT @netspooky: Getting tired of people doing a bunch of llm assisted reverse engineering and making big claims about devices (cough LG and other TVs) and not providing evidence to support these claims. It goes without saying that pretty much all are spying on you, but how it does that matters..

llm逆向方法论 @alkalinesec 原文 ↗

漏洞供给结构变化的一手观察,影响未来可用 exploit 的可得性与价格。

RT @udunadan: @gr4yf0x As I said previously, I had private data points that some smaller actors in exploit business are going out of actions due to AI eating bugs (I semi-ironically refer to explit industry as "a thing of ours" in line with Italian-American lingo for mafia)

ai_agent漏洞市场生态 @0xocdsec 原文 ↗

内核攻击面收敛、检测逻辑迁到用户态,直接影响 EDR 绕过与致盲研究。

Microsoft wants EDR vendors to stop shipping their own kernel drivers. Instead, a single MS driver (wesp.sys) collects kernel events and delivers them to the EDR’s usermode service via a filter communication port. The EDR only writes detection rules https://t.co/Emi5lo6MR1 https://t.co/OfRPPVRSdz

内核edrwindows报告 @papadelospapas_ 原文 ↗

含未达赏金门槛、未上报的漏洞,可作越权与升级链的素材来源。

RT @thedawgyg: Since the @Apple Bug Bounty Program is by far the worst program I've ever hacked on, I am going to be dumping all of the data/POCs/report write ups etc for everything found during the 4 months of hacking them. This will include POCs for vulns that didnt meet their bar for a bounty so wasnt reported to them, so you may be able to find ways to escalate some of them. Ill be uploading the data all this weekend.

移动端poc事件 @cr0nym 原文 ↗

点出 WAF 解析差异仍可利用,以及强 AI 对小团队攻击力的放大。

1. wafoverflow 2. hacktron like securing waf. parser differential ❤️‍🔥 3. php2shell exploitable, but need big brain or frontier ai 4. average attacker small brain, no frontier ai? few exploits? 5. small brain attacker + big brain ai = ?? https://t.co/lOAWBkuD3V

waf绕过llm @S1r1u5_ 原文 ↗

RT @blockscout: Many thanks to @fearsoff for responsibly disclosing a security issue in Blockscout 🐐 The report helped us improve our security, and they’ve been added to our Security GOATS page For open projects, researchers like this are more important than ever 🙏 https://t.co/rQiLS4SdaR

披露开源 @k_firsov 原文 ↗

RT @cyb3rops: I actually expect AI to make software considerably more secure in the coming years. Obvious coding mistakes will become harder to find, and exploiting software vulnerabilities may become much more difficult than it is today. But you don’t need to discover a new software vulnerability to compromise most organizations. Stolen or reused credentials, phishing, session hijacking, social engineering that tricks people into executing malicious code, compromised software suppliers, trojanized applications, malicious updates, insecure configurations, exposed secrets and backups, excessive privileges, abused remote management tools, and weaknesses in authentication and business processes. There are plenty of ways into an organization that don’t require developing an exploit or discovering a new vulnerability. And AI agents can automate and combine these methods just as effectively. They can test different approaches, adapt when something fails, and work their way towards the objective. That’s what this post was about.

llm钓鱼观点 @cyb3rops 原文 ↗

RT @kedamasuzume: ブラウザ完結でAI VTuber・AIキャラクターのライブ配信を行えるWebアプリケーション【AITuber OnAir】の最新版がリリースされたそうです📺 アップデート内容を見ると「会話」:『 GPT-6.1 Terra/Kimi K3.1 Pro 等』 ・「音声」:『 Gemini 3.8 Flash TTS、OpenRouter経由のEleven v4、OpenAIの13種類の声』(本文引用)が追加されたそうです。 最新のモデルですので興味のある方は確認してみて下さい🌟

llm应用 @kernullist 原文 ↗

This week multiple electricronic music instrument companies announced open source hardware synthesizers, a dream for many of us for years. I cannot wait to add my own creative synthesis methods to hardware from @Akai_Pro and @Behringer and whoever else does this.

硬件开源 @Laughing_Mantis 原文 ↗

RT @zoecyber001: When you shut down a modern iPhone, the Power Management Unit (PMU) cuts voltage to the main A-series processor, the OLED display, and the camera array. It leaves a separate Low Power Mode (LPM) rail live on the motherboard. Sitting on that energized rail are three wireless chips: the Bluetooth controller, the NXP NFC controller, and the Apple U1 Ultra-Wideband chip. Apple built this hardware partition to run "Findable After Power Off," Express Transit passes, and Digital Car Keys when your battery dies. Even with iOS completely unloaded from RAM, those three chips stay powered for up to 24 hours on a manual shutdown (and up to 5 hours on automatic battery reserve). Right before iOS powers down, it pre-computes 96 rolling NIST P-224 public keys (one for every 15-minute window over the next 24 hours) and sends them to the Bluetooth chip using Host Controller Interface (HCI) commands. The final HCI command severs communication with the main processor and starts a standalone thread inside the Bluetooth silicon. While the main SoC sits unpowered, the Bluetooth chip wakes up every 2 seconds to broadcast the current P-224 public key over Bluetooth Low Energy. Any passing iPhone or MacBook that hears that broadcast encrypts its own GPS coordinates with your broadcasted public key and relays the packet to Apple servers. In 2022, researchers at the SEEMOO lab at TU Darmstadt audited this low-power hardware chain ("Evil Never Sleeps") and found a major gap between the three wireless chips. While the NFC chip runs encrypted and signed firmware and the U1 Ultra-Wideband chip checks cryptographic signatures at boot, the Broadcom Bluetooth chip has neither encryption nor signature verification on its firmware updates. Broadcom bakes its base Bluetooth firmware into read-only Mask ROM at the factory. To fix bugs and load the standalone Find My feature, iOS writes raw binary patches directly into the chip internal SRAM using a mechanism called PatchRAM (Write_RAM HCI commands). The hardware never checks a cryptographic signature on those RAM patches. If an attacker already has kernel-level code execution on iOS (through spyware or a local jailbreak exploit), they can overwrite the Bluetooth chip PatchRAM in SRAM before the user powers the phone down. When the user drags the power-off slider, iOS halts and wipes main system RAM. Because the PMU keeps the LPM power rail alive, the modified Bluetooth firmware stays in SRAM and keeps executing on reserve battery power. Worse, the TU Darmstadt team showed from iPhone board schematics that the Bluetooth and U1 chips are wired directly to the Secure Element inside the NFC controller via dedicated hardware buses (SE_I2C and R1_WAKE/SYNC) so car keys and transit cards work while the phone is off. Compromised Bluetooth firmware running in LPM can broadcast custom tracking beacons, listen for over-the-air wake commands, or talk directly to the Secure Element over SE_I2C while the main operating system is dead. Apple lets you tap "Temporarily Turn Off Finding" on the shutdown screen and enter your passcode to disable LPM for that power cycle. That toggle works if your phone is clean, because a clean iOS kernel tells the PMU to cut the LPM rail. If your phone is already compromised by kernel spyware, the malware controls the shutdown sequence. It can turn off the screen, play the shutdown animation, ignore your toggle, and tell the PMU to keep the patched Bluetooth chip powered in the background. Once kernel malware can hide inside unverified radio SRAM, a software power-off slider means nothing. Until the battery is physically disconnected from the board or the phone goes inside an RF-shielded Faraday bag, the Bluetooth silicon can stay on the air.

移动端固件研究 @NetAskari 原文 ↗

Kerberos 票据离线爆破覆盖面扩大,域渗透后利用链更完整

RT @sekurlsa_pw: New @hashcat release supports Kerberos etype 3 (DES-CBC-MD5). Newest Windows also supports etypes 19/20 (RFC 8009): 1️⃣9️⃣AES128-CTS-HMAC-SHA256-128 2️⃣0️⃣AES256-CTS-HMAC-SHA384-192 Last number = truncated HMAC length in bits (same idea as the -96 in SHA1-96). #kerberos https://t.co/istzspGUXX

kerberos口令破解工具 @0x64616e 原文 ↗

无需注入 HTML 即可在页面源执行 JS,并附多种 sanitizer 绕过

RT @garethheyes: XSS from a response header 🤯 Firefox will load an XSLT stylesheet from a data: URL in a Link header. Combined with a harmless XML body and you get JS running in the page origin. Plus Safari importing GIFs as SVG and more sanitizer bypasses https://t.co/vJzZ4AvMWI

xss浏览器绕过 @0x64616e 原文 ↗

𝗲𝗠𝘂𝗱𝗵𝗿𝗮 𝗧𝗵𝗿𝗲𝗮𝘁 𝗵𝘂𝗻𝘁𝗶𝗻𝗴 ​ 𝗠𝗲: Do we have a problem? ​ 𝗔𝗱𝘃𝗮𝗻𝗰𝗲𝗱 𝗛𝘂𝗻𝘁𝗶𝗻𝗴: 𝚠𝚎-𝚑𝚊𝚟𝚎-𝚊-𝚙𝚛𝚘𝚋𝚕𝚎𝚖.𝚌𝚘𝚖 ​ 𝗠𝗲: Say no more. 😂 ​ Finally, an IOC with excellent documentation.👍 ​ #CyberSecurity #ThreatHunting #SOC https://t.co/pMjnEfQgcH

威胁狩猎 @0x534c 原文 ↗

RT @SemiAnalysis_: BREAKING NEWS🚨: NVIDIA just submitted its first Rubin NVL72 Kimi K3 submission, and it can reach over 850 tok/s/user, even at 2.8 Trillion parameters, with over 3x higher throughput on @vllm_project! Great work by the NVIDIA team, and excited for Rubin Software 1.0 GA coming soon!

llm推理硬件 @0xocdsec 原文 ↗

RT @OriginalSicksec: @Bugcrowd seems to always find a way to amaze me with its logic - Customer opens message to BC staff - They created a blocker for them - After 7 days they closed my report as N/A due to inactivity 🙃🙃 File new report and wait 2 months for triage again ? No thanks https://t.co/Q8UlTwwvck

漏洞赏金流程 @0xocdsec 原文 ↗

从内存破坏到代码执行的完整链条,可复用的PHP利用工程方法。

RT @AndrewMohawk: Apparently tweeting “I have a working PoC” creates a small obligation to explain how the thing works 😅 So here it is: both PHP bugs, the memory leaks, the ASLR bypass, worker pinning, and how I eventually got to execution https://t.co/iQy3R88cnx

phpaslr利用链poc @moyix 原文 ↗

USB设备侧攻击工程化的完整实现细节,含硬件选型与利用流程。

RT @alexjplaskett: Slides and whitepaper from my @hexacon_fr talk "When Pico's Attack - USB Exploit Engineering" now available at https://t.co/dtVcG37LAt https://t.co/HjUyRKKH8O

议题usb固件工具 @0xocdsec 原文 ↗

RT @jukan05: SK HYNIX: DRAM AND HBM VOLUMES ARE VIRTUALLY SOLD OUT FOR THE NEXT THREE YEARS THROUGH 2029. APPROXIMATELY 50% OF 2027–2028 CAPACITY IS COVERED BY LONG-TERM AGREEMENTS (LTAS), AND THE COMPANY EXPECTS TO MAINTAIN A SIMILAR SHARE GOING FORWARD.

硬件供应链 @0xocdsec 原文 ↗

给出AI智能体挖洞相较传统fuzzer的量化增益与盲区。

RT @natashenka: LLMs are great at finding bugs, but so are fuzzers. @5aelo compared the vulnerabilities found in V8 by BigSleep, an agentic system, with those found by the Fuzzilli fuzzer. https://t.co/KTvm6k5Rhr

ai_agentllm漏洞挖掘浏览器 @0xTriboulet 原文 ↗

RT @layle_ctf: ⚠️A first in Wii emulation! ⚠️ With Starlet LLE and a fresh NAND, we boot through the entire chain: boot0 → boot1 → boot2 → IOS → System Menu → game Then we use my FFCC exploit, The Haxx Bearers, to install BootMii and the Homebrew Channel, all inside my own emulator! https://t.co/k9nqnOQxPe

固件利用模拟器 @alkalinesec 原文 ↗